Design of Keyword Recognition System over Telephone Channel Based on Multi-band Processing

Human speech perception is based on narrow frequency channels, not on whole frequency band. According to the keystone of Human speech perception, the paper introduces multi-band processing into speech recognition system to improve robust. In order to improve the discriminability of speech feature, so as to improve performance of the system, the paper attempts to use independent component analysis (ICA) to transform existing speech feature into another feature space. Experiments over telephone channel show that this system can improve robust to narrow band noises and increase recognition rate obviously.